Why this work keeps a person in the loop
Intelligence analysis starts with source material that is incomplete, contradictory, and sometimes planted on purpose. An analyst validates what is already on file against new field reports, then has to say what the gaps mean. Software can sort and summarize the file. Deciding how much weight a single human source deserves, and how confident a judgment should be, is a different kind of work.
The second reason is accountability. Analysts prepare written assessments, charts, and maps that other people act on, and they have to defend the reasoning behind them to supervisors, prosecutors, and decision-makers. A model can produce a fluent draft. It cannot sit in the room and explain why one interpretation was chosen over another, or carry responsibility when the call turns out wrong.
Third, a lot of the raw material comes from relationships. Interviewing and debriefing sources, and keeping liaison going with other agencies and task forces, is how the useful detail arrives in the first place. Those tasks do not move to software because access, trust, and legal authority sit with the person, not the tool. Robotics barely figures here either: this is desk and network work, so the hardware question that slows automation in trades does not apply.
What AI does, what it helps with, and what stays with analysts
The share of task time that AI can take outright is 6%. That group is routine collection and formatting: pulling and correlating records from intelligence databases, scanning long runs of communications records for patterns, and producing first drafts of reports, charts, and map products. This is the part of the job where speed used to come from long hours. Our coverage method explains how that split is measured.
A similar slice, 77%, is work where AI assists but does not finish. Linking suspects to networks and organizations is a good example: a model can propose connections across case files, and the analyst decides which links are real and which are coincidence. Testing a theory about criminal or terrorist activity works the same way. The tool generates candidate explanations fast; the analyst checks them against sourcing and against what an adversary would want you to believe.
The work that still needs a person is 17% of task time. That is the interviewing and debriefing of sources, the liaison work with partner agencies, and the moment an assessment is briefed and defended. These tasks are small in hours and large in consequence, which is why the headline figure, 67 out of 100 (higher is safer), sits where it does.
What has actually been tested
Not much, directly. Our evidence grade for this job is D, which on our scale means there is no published head-to-head test of AI systems against working intelligence analysts on their own products. Because of that, we publish no parity number for this occupation. We would rather say so than put a figure on an untested claim.
What would settle it is straightforward to describe, if hard to run. You would need a blind comparison on the same raw holdings: analysts and AI systems each produce an assessment, and senior reviewers score the products on sourcing discipline, calibrated confidence, and whether the judgment held up against later ground truth. Deception should be seeded into the inputs, because that is the real test. Until something like that is published and reviewed, the honest position is uncertainty. The quality parity method sets out what each grade requires, and the wider scoring method covers the rest.
The market picture is steadier than the headlines. The Bureau of Labor Statistics puts employment for detectives and criminal investigators, the group this role sits in, at about 114,430 with median pay of $93,790 (BLS, 2025), and projects roughly flat employment of 0.2% from 2025 to 2035.
When the balance could shift
Most likely between 2040 and 2053 (8 in 10 of our scenarios). Two things could pull that earlier. One is accredited deployment of capable models inside classified and law enforcement networks, which is where the data actually lives. The other is cost: the annual software spend we track for this job runs far below what the equivalent human hours cost, and no robotics investment is needed, so once a tool clears security review the business case is easy.
Two things push the other way. Accreditation and data-handling rules move slowly, and a model that cannot touch the holdings cannot do the work. And adversaries adapt: once automated analysis is known to be in the loop, inputs get shaped to fool it, which raises rather than lowers the value of a skeptical reader. For how the window itself is built, see the replacement year method.
How to stay needed as an analyst
Lean into the tasks that sit in the human group. Keep source handling and debriefing sharp, because original collection is the scarce input. Hold the liaison relationships with partner agencies and units, since those shape what you can see. And practice defending judgments out loud, in briefings and in writing, with confidence levels you can justify.
Two skills matter most alongside that. The first is structured analytic technique: hypothesis testing, alternative explanations, and clear confidence language. The second is evaluating model output, which means knowing how to prompt, how to trace a claim back to its source document, and how to catch a confident fabrication before it reaches a product.
What to do: take one recent assessment, redo the collation step with a tool, and spend the hours you save on sourcing checks and the alternative hypothesis.
Nearby work is worth a look if you are weighing a move. Compare this role with Detectives and Criminal Investigators, Private Detectives and Investigators, and Business Intelligence Analysts, which faces a very different task mix. You can also see the whole law enforcement workers family, the government sector, our list of jobs that most need a person, or put two roles side by side with the job comparison tool.